Probability and Statistics Basic concepts II (from a physicist point of view) Benoit CLEMENT – Université J. Fourier / LPSC

Statistics 2 Parametric estimation : give one value to each parameter Interval estimation : derive a interval that probably contains the true value Non-parametric estimation : estimate the full pdf of the population SAMPLE Finite size x i POPULATION f(x; θ ) PHYSICS parameters θ EXPERIMENT INFERENCE Observable

Parametric estimation From a finite sample {x i }  estimating a parameter θ Statistic = a function S = f({x i }) Any statistic can be considered as an estimator of θ To be a good estimator it needs to satisfy : Consistency : limit of the estimator for a infinite sample. Bias : difference between the estimator and the true value Efficiency : speed of convergence Robustness : sensitivity to statistical fluctuations A good estimator should at least be consistent and asymptotically unbiased Efficient / Unbiased / Robust often contradict each other  different choices for different applications 3

Bias and consistency As the sample is a set of realization of random variables (or one vector variable), so is the estimator : it has a mean, a variance,… and a probability density function Bias : Mean value of the estimator unbiased estimator : asymptotically unbiased : Consistency: formally in practice, if asymptotically unbiased 4 biased asymptotically unbiased unbiased

Empirical estimator Sample mean is a good estimator of the population mean  weak law of large numbers : convergent, unbiased Sample variance as an estimator of the population variance : 5 biased, asymptotically unbiased unbiased variance estimator : variance of the estimator (convergence)

Errors on these estimator Uncertainty  Estimator standard deviation Use an estimator of standard deviation : (!!! Biased ) Mean : Variance : Central-Limit theorem  empirical estimators of mean and variance are normally distributed, for large enough samples define 68% confidence intervals 6

Likelihood function Generic function k(x, θ ) x : random variable(s) θ : parameter(s) 7 Probability density function f(x; θ ) = k(x, θ 0 ) ∫ f(x; θ) dx=1 for Bayesian f(x| θ )= f(x; θ ) Likelihood function L ( θ ) = k(u, θ ) ∫ L ( θ ) d θ =??? for Bayesian f( θ |x)= L ( θ )/∫ L ( θ )d θ fix θ= θ 0 (true value) fix x= u (one realization of the random variable) For a sample : n independent realizations of the same variable X

Estimator variance Start from the generic k function, differentiate twice, with respect to θ, the pdf normalization condition: Now differentiating the estimator bias : Finally, using Cauchy-Schwartz inequality Cramer-Rao bound 8

Efficiency For any unbiased estimator of θ, the variance cannot exceed : The efficiency of a convergent estimator, is given by its variance. An efficient estimator reaches the Cramer-Rao bound (at least asymptotically) : Minimal variance estimator MVE will often be biased, asymptotically unbiased 9

Maximum likelihood For a sample of measurements, {x i } The analytical form of the density is known It depends on several unknown parameters θ eg. event counting : Follow a Poisson distribution, with a parameter that depends on the physics : λ i ( θ ) An estimator of the parameters of θ, are the ones that maximize of observing the observed result.  Maximum of the likelihood function rem : system of equations for several parameters rem : often minimize -ln L : simplify expressions 10

Properties of MLE Mostly asymptotic properties : valid for large sample, often assumed in any case for lack of better information Asymptotically unbiased Asymptotically efficient (reaches the CR bound) Asymptotically normally distributed  Multinormal law, with covariance given by generalization of CR Bound : Goodness of fit = The value of is Khi-2 distributed, with ndf = sample size – number of parameters 11 Probability of getting a worse agreement

Least squares Set of measurements (x i, y i ) with uncertainties on y i Theoretical law : y = f(x, θ ) Naïve approach : use regression Reweight each term by the error Maximum likelihood : assume each y i is normally distributed with a mean equal to f(x i, θ ) and a variance equal to Δ y i Then the likelihood is : 12 Least squares or Khi-2 fit is the MLE, for Gaussian errors Generic case with correlations:

Errors on MLE Errors on parameter -> from the covariance matrix For one parameter, 68% interval More generally : Confidence contour are defined by the equation : Values of β for different number of parameters n θ and confidence levels α 13 only one realization of the estimator -> empirical mean of 1 value… n θ  α  1 (0. 5*n θ 2 )

Example : fitting a line For f(x)=ax 14

Example : fitting a line For f(x)=ax+b 15

Example : fitting a line 2 dimensional error contours on a and b 68.3% :  2 = % :  2 = % :  2 =

Example : fitting a line 1 dimensional error from contours on a and b 1d errors (68.3%) are given by the edges of the  2 =1 ellipse : depends on covariance 17

Frequentist vs Bayes Frequentist Estimator of the parameters θ maximises the probability to observe the data. Maximum of the likelihood function 18 Bayesian Bayes theorem links : The probability of θ kwnowing the data : a posteriori to the probability of the data kwnowing θ : likelihood

Confidence interval For a random variable, a confidence interval with confidence level α, is any interval [a,b] such as : Generalization of the concept of uncertainty: interval that contains the true value with a given probability  slightly different concepts For Bayesians : the posterior density is the probability density of the true value. It can be used to derive interval : No such thing for a Frequentist : The interval itself becomes the random variable [a,b] is a realization of [A,B] Independently of θ 19 Probability of finding a realization inside the interval

Confidence interval Mean centered, symetric interval [ μ -a, μ +a] 20 Mean centered, probability symmetric interval : [a, b], Highest Probability Density (HDP) : [a, b]

Confidence Belt To build a frequentist interval for an estimator of θ 1.Make pseudo-experiments for several values of θ and compute he estimator for each (MC sampling of the estimator pdf) 2.For each θ, determine  ( θ ) and  ( θ ) such as : These 2 curves are the confidence belt, for a CL α. 3.Inverse these functions. The interval satisfy: 21 Confidence Belt for Poisson parameter λ estimated with the empirical mean of 3 realizations (68%CL)

Dealing with systematics The variance of the estimator only measure the statistical uncertainty. Often, we will have to deal with some parameters whose values are known with limited precision. Systematic uncertainties The likelihood function becomes : The known parameters ν are nuisance parameters 22

Bayesian inference In Bayesian statistics, nuisance parameters are dealt with by assigning them a prior π( ν ). Usually a multinormal law is used with mean ν 0 and covariance matrix estimated from Δν 0 (+correlation, if needed) The final prior is obtained by marginalization over the nuisance parameters 23

Profile Likelihood 24 No true frequentist way to add systematic effects. Popular method of the day : profiling Deal with nuisance parameters as realization if random variables : extend the likelihood : G(v) is the likelihood of the new parameters (identical to prior) For each value of θ, maximize the likelihood with respect to nuisance : profile likelihood PL( θ ). PL( θ ) has the same statistical asymptotical properties than the regular likelihood

Non parametric estimation Directly estimating the probability density function Likelihood ratio discriminant Separating power of variables Data/MC agreement … Frequency Table : For a sample {x i }, i=1..n 1.Define successive intervals (bins) C k =[a k,a k+1 [ 2.Count the number of events n k in C k Histogram : Graphical representation of the frequency table 25

Histogram 26 N/Z for stable heavy nuclei 1.321, 1.357, 1.392, 1.410, 1.428, 1.446, 1.464, 1.421, 1.438, 1.344, 1.379, 1.413, 1.448, 1.389, 1.366, 1.383, 1.400, 1.416, 1.433, 1.466, 1.500, 1.322, 1.370, 1.387, 1.403, 1.419, 1.451, 1.483, 1.396, 1.428, 1.375, 1.406, 1.421, 1.437, 1.453, 1.468, 1.500, 1.446, 1.363, 1.393, 1.424, 1.439, 1.454, 1.469, 1.484, 1.462, 1.382, 1.411, 1.441, 1.455, 1.470, 1.500, 1.449, 1.400, 1.428, 1.442, 1.457, 1.471, 1.485, 1.514, 1.464, 1.478, 1.416, 1.444, 1.458, 1.472, 1.486, 1.500, 1.465, 1.479, 1.432, 1.459, 1.472, 1.486, 1.513, 1.466, 1.493, 1.421, 1.447, 1.460, 1.473, 1.486, 1.500, 1.526, 1.480, 1.506, 1.435, 1.461, 1.487, 1.500, 1.512, 1.538, 1.493, 1.450, 1.475, 1.500, 1.512, 1.525, 1.550, 1.506, 1.530, 1.487, 1.512, 1.524, 1.536, 1.518, 1.577, 1.554, 1.586, 1.586

Histogram Statistical description : n k are multinomial random variables. with parameters : 27 For a large sample : For small classes (width δ ): So finally : The histogram is an estimator of the probability density Each bin can be described by a Poisson density. The 1 σ error on n k is then :

Kernel density estimators Histogram is a step function -> sometime need smoother estimator On possible solution : Kernel Density Estimator Attribute to each point of the sample a “kernel” function k(u) Triangle kernel : Parabolic kernel : Gaussian kernel : … w = kernel width, similar to bin width of the histogram The pdf estimator is : Rem : for multidimensional pdf : 28

Kernel density estimators If the estimated density is normal, the optimal width is : with n that sample size and d the dimension As for the histogram binning, no generic result : try and see 29

Statistical Tests Statistical tests aim at: Checking the compatibility of a dataset {x i } with a given distribution Checking the compatibility of two datasets {x i }, {y i } : are they issued from the same distribution. Comparing different hypothesis : background vs signal+background In every case : build a statistic that quantify the agreement with the hypothesis convert it into a probability of compatibility/incompatibility : p-value 30

Pearson test Test for binned data : use the Poisson limit of the histogram Sort the sample into k bins C i : n i Compute the probability of this class : p i =∫ Ci f(x)dx The test statistics compare, for each bin the deviation of the observation from the expected mean to the theoretical standard deviation. Then χ 2 follow (asymptotically) a Khi-2 law with k-1 degrees of freedom (1 constraint ∑n i =n) p-value : probability of doing worse, For a “good” agreement χ 2 /(k-1) ~ 1, More precisely (1 σ interval ~ 68%CL) 31 Poisson mean Poisson variance Data

Kolmogorov-Smirnov test Test for unbinned data : compare the sample cumulative density function to the tested one Sample Pdf (ordered sample) The the Kolmogorov statistic is the largest deviation : The test distribution has been computed by Kolmogorov: [0; β ] define a confidence interval for D n β =0.9584/√n for 68.3% CL β =1.3754/√n for 95.4% CL 32

Example Test compatibility with an exponential law : , 0.036, 0.112, 0.115, 0.133, 0.178, 0.189, 0.238, 0.274, 0.323, 0.364, 0.386, 0.406, 0.409, 0.418, 0.421, 0.423, 0.455, 0.459, 0.496, 0.519, 0.522, 0.534, 0.582, 0.606, 0.624, 0.649, 0.687, 0.689, 0.764, 0.768, 0.774, 0.825, 0.843, 0.921, 0.987, 0.992, 1.003, 1.004, 1.015, 1.034, 1.064, 1.112, 1.159, 1.163, 1.208, 1.253, 1.287, 1.317, 1.320, 1.333, 1.412, 1.421, 1.438, 1.574, 1.719, 1.769, 1.830, 1.853, 1.930, 2.041, 2.053, 2.119, 2.146, 2.167, 2.237, 2.243, 2.249, 2.318, 2.325, 2.349, 2.372, 2.465, 2.497, 2.553, 2.562, 2.616, 2.739, 2.851, 3.029, 3.327, 3.335, 3.390, 3.447, 3.473, 3.568, 3.627, 3.718, 3.720, 3.814, 3.854, 3.929, 4.038, 4.065, 4.089, 4.177, 4.357, 4.403, 4.514, 4.771, 4.809, 4.827, 5.086, 5.191, 5.928, 5.952, 5.968, 6.222, 6.556, 6.670, 7.673, 8.071, 8.165, 8.181, 8.383, 8.557, 8.606, 9.032, , D n = p-value = σ : [0, ]

Hypothesis testing Two exclusive hypotheses H 0 and H 1 -which one is the most compatible with data - how incompatible is the other one P(data|H0) vs P(data|H1) Build a statistic, define an interval w - if the observation falls into w : accept H 1 - else accept H 0 Size of the test : how often did you get it right Power of the test : how often do you get it wrong ! Neyman-Pearson lemma : optimal statistic for testing hypothesis is the Likelihood ratio 34

CL b and CL s Two hypothesis, for counting experiment - background only : expect 10 events - signal+ background : expect 15 events You observe 16 events 35 CL b = confidence in the background hypothesis (power of the test) Discovery : 1- CL b < 5.7x10 -7 CL s+b = confidence in the signal+background hypothesis (size of the test) Rejection : CL s+b < 5x10 -2 Test for signal (non standard) CL s = CL s+b /CL b CL b = CL s+b =0.663